Segmentation in medical imaging is crucial for both diagnostic and therapeutic applications. However, it remains a significant challenge under limited annotation conditions. This difficulty is further intensified in multi-modal and heterogeneous clinical environments, where data availability and crossmodal consistency are often constrained. Few-shot segmentation (FSS) methods present a promising solution by learning from a small number of annotated examples. Nevertheless, most existing models assume intra-modal consistency and lack sufficient structural reasoning capabilities, especially in cross-modality scenarios. To overcome these limitations, we propose a coarse-to-fine few-shot segmentation pipeline that integrates a Reference Structure Fusion (RSF) module and a Global Compressed Structural Attention (GCSA) mechanism within a diffusion-based refinement backbone. Specifically, our pipeline first generates an initial coarse segmentation mask from one MRI modality. This coarse mask is then iteratively refined by a denoising diffusion probabilistic model (DDPM)-based segmentation head, guided by fused multi-modal features and structural priors from the multi-stage support-query RSF module. Additionally, we introduce GCSA to ensure global structural consistency by aligning semantic regions between the query and reference images. We evaluate the model on the BraTS2021 brain tumor dataset, which includes four MRI sequences (T1, T1ce, T2, FLAIR), and on other benchmarks to validate its effectiveness. Experimental results demonstrate that our framework achieves significant performance gains, with a Dice score of$88.31\%$and an mIoU of 90. 93% on BraTS2021, outperforming state-of-the-art methods used in experiments by approximately average Dice score of 5%. These improvements validate the robustness and generalization capabilities of our approach across diverse and low-sample segmentation scenarios.

Accurate classification of lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC), the two major subtypes of non-small cell lung cancer (NSCLC), remains challenging due to heterogeneous imaging characteristics and overlapping radiological-pathological patterns. To address this, we propose M2HCNet, a Multimodal and Multiscale Hierarchical Classification Network that integrates computed tomography (CT) and histopathology images for robust lung cancer subtype identification. A multi-scale feature extraction module captures both fine-grained cellular details and global contextual structures, enabling complementary fusion across modalities. The extracted features are then aligned using an optimal transport-based contrastive alignment strategy, which reduces distributional discrepancies and enhances cross-modal consistency under weakly paired conditions. A hierarchical classification mechanism further models inter-class relationships and coarse-to-fine label dependencies to improve discrimination between closely related subtypes. Experiments conducted on the TCGA-LUAD/LUSC cohort demonstrate that M2HCNet achieves 93.8% accuracy, 91.9% F1-score, and 95.5% AUC, outperforming ten recent state-of-the-art baselines. Ablation studies confirm the contributions of multi-scale fusion, cross-modal alignment, and hierarchical learning to overall performance. These findings suggest that M2HCNet effectively leverages heterogeneous yet complementary modalities for reliable lung cancer subtype classification. Future work will extend the framework to patient-matched and multi-institutional datasets to enhance generalization and clinical applicability.

Accurate diagnosis of clinically significant prostate cancer (csPCa) is central to patient management, with the clinical gold standard relying on the interpretation of multi-parametric Magnetic Resonance Imaging (mpMRI) using the PI-RADS v2.1 score [1]. However, this standard is critically challenged by lesion subtlety and substantial interobserver variability, especially within the ambiguous PIRADS 3 category, presenting a significant research gap for automated systems. To address this, we propose a novel twostage deep learning framework designed to systematically separate the tasks of high-recall lesion localization and finegrained PI-RADS classification, while effectively modeling the synergistic relationship between anatomical (T2W) and functional (ADC/DWI) sequences. The first stage introduces a Multi-modal Latent Diffusion Model (LDM) for normative reconstruction and a Learned Difference Sub-network for anomaly scoring, which learns the distribution manifold of healthy tissue to efficiently generate high-recall candidate Regions of Interest (ROIs). Stage 2 employs a dedicated 3D Swin Transformer classifier to perform 5-class PI-RADS prediction on these ROIs. This classifier achieves robust performance by utilizing a Multi-Head Cross-Attention mechanism for dynamic, deep-level feature fusion across T2W and ADC/DWI modalities, and by incorporating prostatic zonal priors to enhance its clinical contextawareness in the final prediction layer. Our framework is rigorously validated on the large-scale, public benchmark PICAI dataset. Quantitative ablation studies confirm that our proposed cross-modal fusion and prior-injection modules are critical drivers of this superior performance, demonstrated across a comprehensive suite of metrics evaluating classification accuracy, clinical utility, and prediction reliability. Our work provides a robust solution that achieves state-of-the-art accuracy and balanced performance, demonstrating strong potential for clinical translation.

Applications of deep learning, in particular Convolutional Neural Networks (CNNs), have shown promise in computer-aided diagnosis, including analysis of osteoarthritis in the knee. Focusing on two of the most popular tasks in medical imaging—segmentation and classification—this work investigates the novelty of adding meta-learning to the multitask learning (MTL) technique for volumetric analysis employing Magnetic Resonance Imaging (MRI) data in the diagnosis of knee osteoarthritis. To enhance the performance of each task, we incorporate recent advances in meta-learning, specifically Model-Agnostic Meta-Learning (MAML) and MetaNet. Experimental results indicate that the innovative integration of meta-learning performs better than all other models. Specifically, MAML compensated for the limited segmentation enhancement seen in the MTL model alone, demonstrating better overall performance. These findings demonstrate MAML’s remarkable capacity to handle challenging multi-task medical image analysis, successfully striking a balance between segmentation and classification accuracy. With the ability to concurrently execute osteoarthritis classification and knee structure segmentation in 3D MRI, this work addresses the computational problems associated with 3D medical imaging and advances the effectiveness of diagnostic models in the field.